[Paper Review] Polyp Segmentation Using Wavelet-Based Cross-Band Integration for Enhanced Boundary Representation

Haesung Oh, Jaesung Lee|arXiv (Cornell University)|Mar 4, 2026

Colorectal Cancer Screening and Detection0 citations

TL;DR

A dual-encoder polyp segmentation model fusing grayscale and RGB features via wavelet-aligned cross-band interactions to improve boundary precision across four benchmarks."

ABSTRACT

Accurate polyp segmentation is essential for early colorectal cancer detection, yet achieving reliable boundary localization remains challenging due to low mucosal contrast, uneven illumination, and color similarity between polyps and surrounding tissue. Conventional methods relying solely on RGB information often struggle to delineate precise boundaries due to weak contrast and ambiguous structures between polyps and surrounding mucosa. To establish a quantitative foundation for this limitation, we analyzed polyp-background contrast in the wavelet domain, revealing that grayscale representations consistently preserve higher boundary contrast than RGB images across all frequency bands. This finding suggests that boundary cues are more distinctly represented in the grayscale domain than in the color domain. Motivated by this finding, we propose a segmentation model that integrates grayscale and RGB representations through complementary frequency-consistent interaction, enhancing boundary precision while preserving structural coherence. Extensive experiments on four benchmark datasets demonstrate that the proposed approach achieves superior boundary precision and robustness compared to conventional models.

Motivation & Objective

• Motivate the need for robust boundary delineation in polyp segmentation under low contrast and illumination variation.
• Investigate boundary cues in the wavelet domain comparing grayscale versus RGB representations.
• Propose a dual-encoder architecture that fuses grayscale and RGB features through frequency-consistent interactions.
• Show that grayscale-based boundary cues can refine RGB structures to improve segmentation accuracy.

Proposed method

• Use two Res2Net-based encoders to extract RGB and grayscale features.
• Introduce a Band-Specific Window Cross-Attention (BS-WCA) module for frequency-aligned cross-modality interaction at corresponding wavelet sub-bands.
• Incorporate a Cascade Dilated Fusion (CDF) block to fuse multi-scale features with dilated convolutions.
• Train and evaluate the model on four datasets (Kvasir-SEG, ClinicDB, ColonDB, ETIS) using Dice and IoU metrics.
• Provide a PyTorch-based implementation setup and reproducibility details.

Figure 1: Structural contrast comparison between RGB and grayscale images in the wavelet domain, showing consistently higher contrast for grayscale across all detail sub-bands.

Experimental results

Research questions

• RQ1 Does integrating grayscale boundary cues with RGB features via wavelet-aligned cross-band interaction improve boundary precision in polyp segmentation?
• RQ2 How does the proposed BS-WCA and CDF design affect boundary accuracy and overall segmentation coherence across datasets?
• RQ3 Are boundary-aware improvements robust to dataset variability (size, illumination, contrast) compared to RGB-only baselines?

Key findings

• The proposed method achieves higher mean Dice and IoU than several baselines across all four datasets.
• Grayscale features provide stronger boundary contrast in the wavelet domain, aiding boundary refinement.
• Frequency-consistent interaction enables high-frequency grayscale details to refine RGB-derived structures.
• The method shows stable performance gains across datasets with varying sizes and imaging conditions.
• Experiments use a dual-encoder architecture with BS-WCA and CDF showcasing improved boundary representation.

Figure 2: Proposed wavelet-based cross-band integration framework that fuses frequency-consistent information from RGB and grayscale features for enhanced boundary representation.